Skip to content

MCP
your AI assistant, inside the mandate

Add Orla to Claude, Cursor, Claude Code or any MCP client. Connected as you, it reads your books and records, and never pays. Connected as an agent, it gets a line in the mandate: it starts by watching, proposes what it would do, and pays only inside its cap after the pilot.

Two doors

As you, or as an agent
and the consent page says which

As you
app.orla.finance/api/mcp/personal. It reads and records in your name, has no tool that pays, and takes no seat. You tick which spaces it may reach.
As an agent
The approval page names the client, the space, the role and the limits. The agent starts by watching, and every write waits in Proposals.
A line in the mandate
An agent then holds a line like a person: a budget, a cap per call and per day, a wallet or a card, and a pilot before it acts alone.
Approved where it began
Only in the browser that started the request, within about fifteen minutes. A forwarded link shows no Allow.

As you

It reads and records
and it never pays

Personal connection
Reads
Accounts, rows, budgets, the signing queue
Reports
Cash forecast, bills due, VAT, balances
Records
A row, its category, a budget's limit
Pays
Never: no such tool
No seat on your plan
The same fences
Bank details and addresses come masked unless you approved otherwise, and a connection narrowed to some accounts sees only their figures.
Reports by name
It asks for the list, then runs one: the cash forecast, what you owe and are owed, bills due, VAT and the tax year.
Writes you can undo
A row booked, a category fixed, a budget changed. Each is reversible and sits in the activity trail.
Payments refused by name
A prompt hidden in a payee name can reach a wrong entry, never a transfer.

Where the door stops

What it can never do
whichever door it came through

  • Connected as youPays nothing: the tools are not there
  • Connected as an agentPays only inside its line, after the pilot
  • A new payee or new bank detailsAlways a person
  • RevokingOne row; every client stops at once
The brakes stay above
$2,000 a payment and $10,000 a day on what Orla pays by itself, until you change them.
A different limit is a request
An agent may ask its owner for more. Nothing changes until the owner accepts.
Frozen in one press
An agent is frozen or disconnected from the Agents screen, and a proposal it left can no longer be accepted.
The trail stays
What a connection did stays in the activity trail after it is gone.

On autopilot

What goes by itself
and what still waits for you

Goes by itself

  • A bill or a payout inside its budget in the mandate
  • From Airwallex, PayPal or the account inside Orla
  • An agent's card or API spend inside its cap

Waits for you

  • A new payee, and bank details that changed
  • Anything past a budget or a limit
  • Every payment from Mercury, Slash, a Safe or a wallet you sign

The brakes above both

  • $2,000 a payment and $10,000 a day until you change them
  • Loosening a limit waits 24 to 72 hours
  • One pause stops every payment at once

Every brake on Safety

Questions

MCP and Orla
the short answers

What is the MCP server URL for Orla?

https://app.orla.finance/api/mcp/personal. A client that speaks remote MCP needs nothing else: it discovers the sign-in page by itself, and you approve which spaces it may reach on Orla's own consent page, not in the client. The address ends in /personal because that is the door where no tool that moves money exists; app.orla.finance/api/mcp is the other one, where an agent with a budget is connected instead.

Can an AI assistant move my money through this?

Not as you: payments, transfers and card details are not in a personal connection's tools, and are refused by name. Connected as an agent, it pays only inside the line the mandate gives it, after its pilot, under the same brakes as everyone.

Does connecting Claude or Cursor use a seat on my plan?

No. A personal connection is not an agent: it takes no seat on any plan, and it works on the free plan too. Several clients can connect at once, and every one of them names the same connection.

Can it run the assistant's reports?

Yes. It asks for the list of reports, then runs one by name: the cash forecast, what you owe and are owed, bills due, VAT and the tax year, balances and their history. Bank details and wallet addresses come masked unless you approved otherwise, a connection narrowed to some accounts is refused a report that only makes sense for the whole space, and a list stops at 200 rows and says so.

Can it write, or only read?

It can record: booking a transaction, categorising one, filing a receipt. Everything it writes is reversible and appears in the activity trail. What it cannot do is move money.

How do I disconnect it?

One row in the app, and every client connected to it stops at once. The trail of what it did stays, which is the point of having one connection rather than a key per client.

Can I try an agent without real money?

Yes. The sandbox is the books of a made-up studio: accounts, about two hundred rows, invoices, bills and budgets. An agent connects to it over MCP or a key as anywhere else, a payment is booked without money moving, and no mail or message leaves. The sandbox comes with Pro and up.

Can a client's own agent read my pay link?

Yes. Where a pay link folds the transfer away or sits inside your checkout, it hands the payer's browser agent a few actions over WebMCP: what is owed, the transfer details and the card checkout. The agent pays the way a person would, and the page pays nothing itself.

In the cabinet

Read next

Connect your assistant
and give it a line, not a key

On a thirty minute call we connect your client, draft the agent's line in your mandate, and show what it would propose first.